Model accuracy
We don't hide the model behind "5000 simulations". Here's how often our pre-match prediction is right - and where it keeps missing. Bookmaker odds are the ceiling we approach but do not beat.
Our live track record
86 predictions settled, from 13/06/2026 to 07/07/2026. Each logged BEFORE the match, compared to the real result.
This is still a small sample - read the numbers as an early signal, not a verdict. They grow every round.
Favourite accuracy
62%
62 out of 100
How often the most likely outcome (1, X or 2) actually happened - across all matches, draws included.
On decided matches (68)
78%
78 out of 100
Same thing, but draws excluded from the denominator. Higher, because a draw can't be the single most likely pick.
Brier / RPS
0.517 / 0.180
Quality of the whole probability distribution (lower = better). Naive baseline on the same set: 0.632 / 0.233.
Calibration
0.096
When the model says "60%", how often it's really 60%. Mean error: 0.096 (closer to zero = more honest).
Our blind spot: draws
The model picked a draw as most likely in 0 of 100 matches - yet draws happened in 21 of 100. Not a fluke: strength-based models structurally under-rate draws. That's why on an even match we say "toss-up" rather than name a favourite.
Data as of 07/07/2026.
What to expect
- -Bookmaker odds are the ceiling for prediction quality. In the Bundesliga the odds favourite is right ~52.5% (~69.9% on decided matches) - we sit ~3.8pp below that ceiling. We do NOT claim to beat bookmakers.
- -Draws are a structural blind spot for any such model: ~26% of results, yet models pick them as most likely only a few percent of the time.
- -Predictability depends on the league - from ~50% in the most even competitions to ~64% in the most predictable.
- -We show probability as natural frequencies ("68 out of 100 simulations") because it reads more clearly than a raw percentage.
How we measure it
Before each match we log the full prediction vector (win, draw, loss chance) from a Monte Carlo simulation. After the real whistle we compare it to the result. We compute favourite accuracy, Brier and RPS (distribution quality) and calibration - the same metrics Opta or the old FiveThirtyEight used.
A separate, earlier backtest of the raw Elo rating gave 66.1% favourite accuracy on decided matches (~49.6% in the 1/X/2 convention with draws). That's a different measurement from the live track record above - we give it as a reference point, not our current number.
See the model in action
Set a lineup and tactics, and the simulation shows whether your plan would have worked.